AGI-26 · San Francisco · Research Paper

Is AI
conscious?
Now we can measure it.

For the first time, functional consciousness has a number.

Everyone asks whether AI is conscious. Almost nobody asks how to measure it. This paper proposes a computationally tractable metric — the Functional Consciousness Score — grounded in information theory and benchmarked on real systems, from a Waymo self-driving taxi to the human mind. The results are both surprising and reassuring.

FCS = R · P = B · D̄ · P Representational Capacity × Reasoning Power
Read the Paper See the Scores ↓
Explore
The Score

How conscious
is your AI system?

The Functional Consciousness Score measures a system's observable capacity to access and reason about its own internal states — not what it "feels like" to be that system, but what it can actually know about itself.

The results reveal a hierarchy with a clear message: current AI systems, however powerful, are operating with a fraction of human self-awareness. The gap is not philosophical — it is numerical.

FCS = R · P, where R (Representational Capacity) measures how richly a system models its own states, and P (Reasoning Power) measures how effectively it can reason over those models.

A map has rich spatial data but zero reasoning — it scores zero. A stateless language model has immense reasoning power but no persistent self-model — it also scores zero. Functional consciousness requires both.

The multiplicative structure is intentional: either dimension alone is insufficient. Only systems that combine rich self-representation with powerful inference achieve meaningful FC scores.

Reasoning vs Capacity Diagram
System B (variables) D̄ (bits/var) P (reasoning) FCS Scale
MapStatic data · no reasoning
 ~1,000 ~40 0 0
Stateless LLMTransformer · no persistent state
 0 0 ~3,300 0
LIDACognitive architecture · symbolic
 ~20 ~4 ~33 ~2,600
Roomba + SLAMSpatial self-model · limited reasoning
 ~18 ~8 ~39 ~5,600
ACT-RProduction system · bottlenecked
 ~20 ~8 ~50 ~8,000
Waymo L4Autonomous vehicle · kinematic domain
 ~40 ~14 ~133 ~74,500
Generative AgentsLLM + memory + reflection · episodic
 ~130 ~100 ~497 ~6.5M
Human (kinematic)Biological · cerebellar forward model
 ~550 ~10 ~1,826 ~10M
Human (working mem.)Biological · reflective reasoning
 ~330 ~14 ~3,000 ~13.9M
0
Both extremes score zero
A map with rich data but no reasoning scores zero. A stateless LLM with immense reasoning but no self-model also scores zero. FC requires both dimensions simultaneously.
×87
The agentic leap
Adding memory, reflection, and persistent state to a stateless LLM — as in Stanford's Generative Agents — multiplies the FCS by 87×. The scaffold, not the model, is where functional consciousness lives.
188×
The human gap
The most capable current AI agent scores roughly 188 times lower than the human working memory baseline. This gap is not opinion — it is arithmetic. And it should be reassuring.
Cognitive Shape

Where does a system
know itself?

A single FCS number tells you how much functional consciousness a system has. The cognitive shape tells you where.

We identified 46 distinct self-models across ten functional domains — from body awareness and spatial reasoning to social understanding, ethical self-monitoring, and meta-reflection. Plotting a system's coverage against the human baseline reveals its unique cognitive fingerprint.

The ten domains were derived from a bottom-up analysis of Virginia Woolf's stream-of-consciousness prose — a dataset uniquely dense in first-person self-reference — using a methodology called Functional Self-Model Analysis (FSMA).

FSMA asks: what internal models must a system possess to consistently produce a given output? If a system reliably describes its own emotional state, it must functionally model that state — regardless of whether it "feels" anything.

Body Spatial Action Goal Cognitive Information Emotional Social Meta Ethics
Waymo L4
VectorNet / MPC / MCTS
Body Spatial Action Goal Cognitive Info
74,500
FC Points
LIDA
Conscious Cycle (GWT)
Body Spatial Action Goal Cognitive Info
~2,600
FC Points
Generative Agents
LLM + Reflection Loop
Body Spatial Action Goal Cognitive Info
~6.5M
FC Points
Human
Biological · working memory
Body Spatial Action Goal Cognitive Info
~13.9M
FC Points · baseline
▶
Narrow but deep
Waymo's shape is concentrated in Body and Spatial domains — it knows its own position and trajectory exquisitely well. Everything else is absent. This is a system that is highly conscious of where it is, and entirely unconscious of what it wants, feels, or knows.
▶
Asymmetric and social
Generative Agents have almost no body awareness but rich cognitive, social, and goal self-models. They live entirely in the mind, with no body. This asymmetry reveals both the power and the fragility of purely linguistic agents.
▶
Full-spectrum awareness
The human baseline fills all ten domains. This breadth, combined with high reasoning power, produces the exponential FCS advantage. No current AI system approaches this coverage — and the gap is largest precisely in the domains most people fear AI is replicating.